Interpretable whole-brain prediction analysis with GraphNet